CSCE790T Medical Image Processing

1
CSCE790T Medical Image Processing
University of South Carolina Department of
Computer Science
3D Active Shape Models Integrating Robust Edge
Identification and Statistical Shape Models
2
Overview

• Introduction
• Motivation
• General ASM Algorithm
• Robust Edge Detection
• Unified Cost Function
• Experiments / Results
• Conclusion

3
Introduction

• Collaboration with UNC departments of computer
  science, and psychiatry
• Submitted to MICCAI 07
• Propose two new strategies to improve 3D ASM
  performance
• Developing a robust edge-identification algorithm
  to reduce the risk of detecting false edges
• Integrating the edge-fitting error and
  statistical shape model defined by a PDM into a
  unified cost function

4
Introduction

• Apply the proposed ASM to the challenging tasks
  of detecting the left hippocampus and caudate
  surfaces from an subset of 3D pediatric MR images
• Compare its performance with a recently reported
  atlas based method.

5
Overview

• Introduction
• Motivation
• General ASM Algorithm
• Robust Edge Detection
• Unified Cost Function
• Experiments / Results
• Conclusion

6
Motivation

• Segmentation facilitates the discovery of
  diseased structures in medical images
• Two neurological shape structures of interest
• Caudate Nucleus
• body movement and coordination
• cauda (tail)
• Hippocampus
• memory and coordination
• hippo (horse) and Kampi (curve)

7
Motivation
http//www.emedicine.com/radio/topic443.htmtarget
2
8
Motivation
http//www.sci.uidaho.edu/med532/basal.htm
9
Motivation

• Hippocampus, and Caudate related to the following
  areas of research
• Epileptic seizures (MTS)
• Alzheimer disease
• Amnesic syndromes
• Schizophrenia
• Parkinson's disease
• Huntington's disease

10
Overview

• Introduction
• Motivation
• General ASM Algorithm
• Robust Edge Detection
• Unified Cost Function
• Experiments / Results
• Conclusion

11
General ASM Algorithm

• Initial placement of point distribution model
  (PDM) mean shape inside image volume T (v s, t,
  ? )
• Generate gradient magnitude values for each voxel
  location in 3D image volume
• while not(convergence)
• Identify strongest edge for each landmark point
  along its search path
• Using this edge information determine new ASM
  shape
• Update PDM global transform T(s, t, ?) and local
  transform variables
• Verify new ASM shape with PDM shape space limits
• If global, and local transform variables are not
  longer changing ASM has converged

12
Overview

• Introduction
• Motivation
• General ASM Algorithm
• Robust Edge Detection
• Unified Cost Function
• Experiments / Results
• Conclusion

13
Robust Edge Detection

• Identify boundary edges of desired surface
  structure inside image volume
• Each edge is represented by an gradient magnitude
  value
• Stronger edges have larger gradient magnitude
  values

14
Robust Edge Detection
Example sagittal plane edges for hippocampus
Image slice
Gradient magnitude slice
15
Robust Edge Detection
Example coronal plane edges for hippocampus
Image slice
Gradient magnitude slice
16
Robust Edge Detection

• Boundary edges are identified along search paths
  for each landmark point
• Search paths are defined by profile locations (?)
  along each landmark points normal vector

17
Robust Edge Detection

• Additionally, each landmark points normal vector
  is determined by the surface mesh

5
2
A
n6 ¼ x (nD nE nF nG) n6/ n6
1
E
F
B
6
C
8
D
3
G
4
7
18
Robust Edge Detection
19
Robust Edge Detection

• Generally, edges detection along search paths are
  considered dangerous
• Subject to noise
• Spurious (false) edges

20
Robust Edge Detection

• Propose an new neighborhood solution
• Spatially consistent profile location (?i?)
• Reduces the likelihood of an false edge

21
Overview

• Introduction
• Motivation
• General ASM Algorithm
• Robust Edge Detection
• Unified Cost Function
• Experiments / Results
• Conclusion

22
Unified Cost Function

• Traditionally each of the models local transform
  variables (bi) are updated after the ASM shape
  is found
• If the ASM shape (u) is not defined within the
  limits of the PDM shape space the local transform
  variables (bi) are rescaled appropriately
• Shape information may be lost
• Re-active solution

23
Unified Cost Function

• Steps in shape deformation where ASM shape not
  within PDM shape space limits

24
Unified Cost Function

• Proposed solution implemented by an unified cost
  function
• Pro-active solution
• Efficiently solved as an quadratic programming
  problem

25
Unified Cost Function

• The cost function can be viewed as,

• vT (3nx1) vector global transformed mean
  shape
• DT -1 (3nx3n) matrix global transformed
  inverse covariance matrix
• u (3nx1) vector initial PDM mean shape or
  previous ASM shape
• N (3nxn) matrix the normal vectors
• ? (nx1) vector profile locations of the
  most stable edges
• ? (nx1) vector most optimal profile
  locations

26
Overview

• Abstract
• Motivation
• General ASM Algorithm
• Robust Edge Detection
• Unified Cost Function
• Experiments Results
• Conclusion

27
Experiments Results

• Developed using ITK and VXL C open source
  libraries
• Subset of 10 high resolution MRI brain images
  from pediatric study
• 256x256x192 resolution
• Inter-voxel spacing 1.0mm

28
Experiments Results

• Left hippocampus PDM
• 42 shape instances
• 642 corresponded landmark points
• Corresponded using MDL
• Left caudate nucleus PDM
• 85 shape instances
• 742 corresponded landmark points
• Corresponded using SPHARM

29
Experiments Results

• Each PDM mean shape was manually initialized
  using Insight-SNAP
• Convergence was achieved when either the global
  transform variables or mahalanobis distance
  between ASM shape and PDM mean shape were at an
  minimum.
• Convergence was typically achieved between 5 to 7
  ASM iterations using /- 4 (k9) profile
  locations along each landmark points normal vector

30
Experiments Results
31
Experiments Results
32
Experiments Results

• ASM segmented performance was compared against
  Atlas-based method
• Performance was evaluated using the following
  measures
• Pearson correlation coefficient volumetric
  correlation
• Dice coefficient volumetric overlap

33
Experiments Results
34
Overview

• Abstract
• Motivation
• General ASM Algorithm
• Robust Edge Detection
• Unified Cost Function
• Experiments / Results
• Conclusion

35
Conclusion

• Presented two new strategies to address
  limitations of current ASM.
• Robust edge detection to reduce likelihood of
  spurious edge
• Pro-active solution ensure ASM approximated shape
  is defined within PDM shape space limits using
  unified cost function
• Additional research is required to address the
  sensitivity of the initial placement
• Implement fully-automatic method